MIFormer: An Effective Semantic SegmentationNetwork for Foggy Weather conditions

Reliable operation is a key metric for autonomous driving, but the factors related to the environment such as weather and lighting can impair the perception capability of autonomous vehicles. Unsupervised adaptation in foggy circumstances is difficult due to the significant fluctuations in the scene's visibility caused by weather conditions, including smog, fog, and haze. This paper presents an effective network using multi-teacher intermediate domain adaptation (MIFormer), which offers a fresh approach to adapting semantic segmentation for the purpose of handling dense foggy scenes. This network integrates training from an intermediate domain to reduce the domain gap caused by heavy fog weather. To further increase segmentation accuracy, we propose three training strategies: Rare Class Sampling, Multi-teacher knowledge distillation, and Mixup improved by Uncertainty principle. We assess the performance of the suggested model across scenarios involving the adaptation from actual clear-weather scenes to real foggy scenes, as well as from synthetic non-foggy images to real foggy scenes. The results have demonstrated that MIFormer’s performance in semantic segmentation with foggy images is better than that with ResNet-38, BDL, CBST, MLST and FogAdapt. In particular, when adhering to the standard configurations and comparing against some mainstream methods, MIFormer achieves a segmentation performance of 69.1 mIoU adapted from GTA to foggy-Cityscapes and 61.8 mIoU adapted from Synthia to foggy-Cityscapes. Additionally, it can even segment rare classes effectively, such as train, bus, and truck.